Attention Mechanism For Scene Text Recognition

Text is the main tool for information transmission,and it exists widely in various scenes.Because the correct recognition of the text in the scene helps to understand the information in the scene,it is always a key research topic for scholars to use algorithms to allow the machine to correctly recognize the text information from the picture.Unlike the document text recognition task,the main challenge of the scene text recognition task is that the complex scene environment leads to the diversity of text pictures,such as the natural light is too strong,which causes the reflection of key parts of certain text,or the brightness is too low,so that the outline of the text is not obvious;Or because of the problem of shooting angle,the text in the picture shows a certain deformation,which will cause the text shape to become irregular;the above factors will cause difficulties in text recognition.In recent years,many research results have shown that the attention mechanism algorithm based on deep learning has diversity and flexibility in processing sequence problems,so this paper studies the attention algorithm to solve the problem of scene text recognition algorithm.The research work and contributions of this article mainly include:1.We improves the encoding network for extracting image features,which alleviates the effect of differences in the objective distribution of training data and test data,and makes the scene text recognition model more robust.2.Regarding the recognition of irregular text,this paper improves on the basis of the classic one-dimensional attention mechanism,and proposes two scene text recognition methods based on two-dimensional attention mechanism.The method designed in this paper makes full use of the context information of the character currently being recognized to generate a twodimensional feature attention distribution.This attention distribution locates the location of the local feature of the corresponding character in the two-dimensional feature map.This way of obtaining two-dimensional attention distribution does not require the use of character-level annotation information,which reduces the cost of training.3.In order to better improve the attention mechanism,this paper proposes a scene text recognition algorithm based on the cross-attention mechanism,which allocates the process of extracting the corresponding characters from the two-dimensional features to two stages of vertical and horizontal directions,thus The network has the advantages of two-dimensional attention mechanism and can flexibly improve the decoding process of its attention network.Experimental results show that the two proposed scene text recognition algorithms based on two-dimensional attention mechanism have a significant improvement in performance compared with the traditional scene text recognition algorithms based on one-dimensional attention mechanism.After the scene text recognition algorithm based on the cross-attention mechanism introduces the attention interaction mechanism in the field of natural language processing,its recognition effect has also been significantly improved.